Machine learning-based integration develops a hypoxia-derived signature for improving outcomes in glioma.

The growth of glioma is frequently accompanied by a hypoxic microenvironment. Nevertheless, the clinical implications of hypoxia have not been extensively investigated. Single-cell RNA sequencing analysis indicated a heterogeneous communication between different types of cells in the hypoxic microenvironment. Two hypoxia-related glioma subtypes, C1 and C2, show distinct prognostic and molecular differences. Subtype C2 gliomas have more immune and stromal cells, higher immune checkpoint gene expression, and worse prognosis than those in C1. Using machine learning, we developed an 11-gene signature predicting clinical outcomes in six cohorts, validated by RT-qPCR, effectively distinguishing high-risk and low-risk patients and reliably predicting overall and relapse-free survival. Moreover, the risk score is more accurate than conventional clinical variables, molecular characteristics, and 100 previously published signatures. High-risk gliomas show increased CD163, PD1, HIF1A, and PD-L1 expression. We developed a hypoxia-related classification to guide treatment decisions and a reliable prognostic tool.